KAIST Humanoid v0.9 pushes 35 kg: what the video shows

In a video published on 19 October, KAIST's Dynamic Robot Control & Design Laboratory shows a humanoid walking while moving blocks into the letters “KAIST”. The team reports two separate results for Humanoid v0.9: pushing objects weighing up to 35 kg and lifting objects weighing up to 20 kg. These are demonstrations under particular conditions, not a certified payload rating.

What the demonstration shows

The robot approaches and moves the blocks while walking. KAIST credits RAILAB with developing the arms and describes onboard visual feedback. The research concerns *whole-body loco-manipulation*: coordinating the legs, torso and arms when contact with an object changes balance.

Pushing a 35 kg block is different from carrying 35 kg. The floor supports a pushed object's weight, and friction, surface and shape affect the force required. The separate 20 kg figure concerns lifting, according to the researchers. Combining the two figures would overstate the result.

Why it matters, and what remains open

Pushing transmits force through the arms and torso to the legs. The robot must stay balanced while continuing toward its target. The video illustrates that coordination and the use of onboard perception to locate the blocks.

KAIST describes skills developed through reinforcement learning and human demonstrations. The public material does not include a complete log of attempts, standardized friction conditions, possible operator interventions or independent measurement of the loads. It does not establish reliable performance with unfamiliar objects or over hours of work. The filming date is undocumented; 19 October is the video's publication date.

Humanoid v0.9 is a research platform, not a home product with a published price or support plan. A buyer should ask what was moved, under which conditions, how often the task succeeded and what human help was involved. The ROBOLERO robot catalogue contains model profiles; the KAIST test illustrates a research direction.

How to read a strength demonstration

An object's mass alone does not tell us how much force the robot applied. During a push, the surface, friction, slope and direction of travel all matter. During a lift, the robot must support the object while controlling its own balance. That is why 35 kg pushed and 20 kg lifted are not two versions of the same payload figure. The public material does not say whether the lab repeated the tasks on different floors or how close the robot was to losing stability.

The video is a selected sequence. It helps explain the coordination of walking and manipulation, but it cannot establish a success rate. A successful attempt may depend on prepared conditions such as the blocks' starting positions, a known surface and a defined route. That is normal for a research demonstration. It becomes misleading only when presented as proof that the robot can perform any real-world job. Comparisons with other humanoids require the same caution: without common objects, surfaces and success criteria, the larger number is not automatically the more reliable robot.

What would make the result more useful

Repeated trials with the same loads on different surfaces would show how well the behavior transfers. A public test record could report success rates, human interventions, completion times and the conditions under which the robot stops or fails. It would also separate the instruction given before a task from any commands or corrections during execution. Such a protocol would let readers compare the same task across labs instead of comparing edited clips.

Those measurements are not in the materials reviewed here, so we do not credit KAIST with results it has not documented. What the released sequence does show is meaningful: the robot combines walking, onboard perception and contact with heavy objects in the task shown. How often and how long it can do that in unfamiliar environments remains an open question. This is the distinction to carry over when assessing a robot offered to buyers: a vivid demonstration can identify a promising capability, while reliability needs repeatable evidence.

Sources

• KAIST DRCD Lab, original demonstration video.
• Humanoids Daily, report on the demonstration.
• Interesting Engineering, analysis of the video and reported loads.